Merge pull request #30 from RICHARDNAN/csx-main-fix

删除废弃代码

Author: cen121212

doc/zh/DeepseekR1_tutorial_zh_for_Ascend_NPU.md

@@ -37,8 +37,9 @@ conda install -c conda-forge libstdcxx-ng  # 安装`GLIBCXX-3.4.32`
 apt install zlib1g-dev libtbb-dev libssl-dev libaio-dev libcurl4-openssl-dev
 pip3 install numpy==1.26.4  # 适配torch/torch_npu
 pip3 install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cpu
-pip3 install packaging ninja transformers==4.43.2 fire protobuf attrs decorator cloudpickle ml-dtypes scipy tornado absl-py psutil
+pip3 install packaging ninja fire protobuf attrs decorator cloudpickle ml-dtypes scipy tornado absl-py psutil
 pip3 install sqlalchemy
+pip3 install transformers==4.57.1 #此处注意运行时transformers版本要求4.57.1(其他版本未验证)
 #pip3 install cpufeature  # only for x86
 ```
 
@@ -122,7 +123,7 @@ python ktransformers/server/main.py \
 --gguf_path /mnt/data/models/DeepSeek-R1-q4km-w8a8 \
 --model_name DeepSeekV3ForCausalLM \
 --cpu_infer 60 \
---optimize_config_path  /home/huawei/ktransformers/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-300IA2-npu-serve.yaml \
+--optimize_config_path  ./ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-300IA2-npu-serve.yaml \
 --max_new_tokens 128 \
 --max_batch_size 4 \
 --use_cuda_graph \

ktransformers/models/ascend/custom_ascend_modeling_deepseek_v3.py

@@ -114,16 +114,6 @@ def print_callback(self, param):
             print("########################################")
             print("hidden_states is ", hidden_states)
             print("########################################")
-        # with torch.npu.stream(self.call_stream):
-        #     position_ids, page_idx, page_offset, block_tables, hidden_states, bsz, q_len, hidden_size = param
-        #     print("########################################")
-        #     print("position_ids is ", position_ids)
-        #     print("page_idx is ", page_idx)
-        #     print("page_offset is ", page_offset)
-        #     print("block_tables is ", block_tables)
-        #     print("hidden_states is ", hidden_states)
-        #     print("#########################################")
-
 
     def forward(
         self,
@@ -172,27 +162,8 @@ def forward(
             q_len_raw = None
             kv_len_raw = batch.minibatch.d_kv_len_list
             bsz_real = None
-            # if utils._USE_NPU_GRAPH:
-            #     from libgraph_capture import graph_capture_launch_callback
-            #     param = (position_ids, page_idx, page_offset, block_tables, hidden_states, bsz, q_len, hidden_size)
-            #     graph_capture_launch_callback(self.print_callback, param, 1, self.stream.npu_stream)
-            # else:
-            #     param = (position_ids, page_idx, page_offset, block_tables, hidden_states, bsz, q_len, hidden_size)
-            #     self.print_callback(param)
-
 
-        # with torch_npu.npu.stream(self.stream):
-        # print_ex("####: before decode layer...")
         for i, decode_layer in enumerate(self.model.layers):
-            # if not is_prefill:
-            #     if utils._USE_NPU_GRAPH:
-            #         from libgraph_capture import graph_capture_launch_callback
-            #         param = (hidden_states, )
-            #         graph_capture_launch_callback(self.print_callback, param, 1, self.stream.npu_stream)
-            #     else:
-            #         param = (hidden_states, )
-            #         self.print_callback(param)
-            # attn
             residual = hidden_states
             hidden_states = decode_layer.input_layernorm(hidden_states)
 

ktransformers/server/backend/interfaces/ktransformers.py

@@ -232,12 +232,6 @@ def prefill(self, input_ids: torch.Tensor, is_new: bool, temperature: Optional[f
                 )
                 self.seq_length = 1
 
-            # flat_prev_ids = self.generated_ids.flatten()
-            # for i in range(min(self.seq_length, flat_input_ids.shape[0]) - 1):
-            #     if flat_input_ids[i] == flat_prev_ids[i]:
-            #         same_prefix += 1
-            #     else:
-            #         break
 
             logger.debug(f"same prefix len: {same_prefix}")
             self.cache.remove_suffix(same_prefix)

ktransformers/server/utils/serve_profiling.py

@@ -106,19 +106,3 @@ def get_stat(self):
 
 PROF_TIME_STAT = ProfTimeStat()
 
-
-# j=0
-# start_time = PROF_TIME_STAT.record_start_time()
-# for i in range(500):
-#     j+=1
-# PROF_TIME_STAT.add_time_stat(ProfStatKey.ExpertsSummitCurrLayer, start_time, False)
-
-# for i in range(500):
-#     j+=1
-# PROF_TIME_STAT.add_time_stat(ProfStatKey.ExpertsSummitCurrLayer, start_time, False)
-
-# for i in range(500):
-#     j+=1
-# PROF_TIME_STAT.add_time_stat(ProfStatKey.ExpertsSummitCurrLayer, start_time, False)
-
-# PROF_TIME_STAT.print_all()